The Research And Improvement Of Speech Enhancement Algorithms

Speech enhancement, which is an important branch of speech signal processing , is widely used in various practical speech processing system. In practical applications, the speech signal to be analyzed is usually contaminated by the environment noise which reduces the performance of speech processing systems. In order to improve the signal processing ability of the speech processing system, it is necessary for speech enhancement to be used to suppress the noise and restore the pure speech.First, the paper introduces the basic theory of speech enhancement technology briefly, and several conventional algorithms and the evaluation methods are reviewed. Then the speech enhancement based on the wavelet technology and the spectral subtraction algorithm is analyzed in detail.The wavelet is a flexible time-frequency analysis technology, which is a powerful tool for the non-stationary signal like the speech. The paper introduces the basic principle of the wavelet and the wavelet packet and then discusses the threshold shrinkage algorithm. Although the threshold shrinkage algorithm is simple and of satisfying performance, but there still has an important issue that is the problem of the over-thresholding, which means not only the unwanted noise is suppressed but some speech segments. Consequently the perceptual quality of the enhancement speech is degraded. In order to overcome the problem, a improved wavelet denoise threshold is proposed, which decomposes the noisy speech by the perceptual wavelet packet and then the denoise threshold is adjusted by the speech-present probability which is obtained by tracking the energy of the node coefficients. The experiment result based on the Matlab shows that the proposed denoise threshold can avoid the over-thresholding phenomenon well and the enhanced speech has the better hearing quality.Furthermore, the spectral subtraction is also analyzed in this paper. The spectral subtraction speech enhancement is utilized broadly because it is simple and easy for the real-time processing, but the estimation of the noise power spectrum is calculated by the smoothing noisy speech power spectrum during the speech-absent segment, witch is not accurate and degrades the quality of the enhanced speech. In order to improve the denoise ability of the spectral subtraction especially under the non-stationary noise environment, an improved spectral subtraction algorithm is proposed. When caculating the estimation of the noise power spectrum, the spectrum variance endpoint detection is used first and if the noisy speech is speech-absent, then the noise power spectrum is upgraded for all frequency bins, or the power spectrum is divided to several subbands according to the Bark frequencies and tracks the noise by the improved minima controlled recursive averaging method in every individual band. The improved noise estimation approach can estimate the noise power spectrum more accurately while reduce the computation because of considering the characteristic of the speech spectrum and the hearing perception. Furthermore, a new adaptive subtraction factor and spectral floor parameter is used in the improved spectral subtraction, which can reduce the noise better. And at last the effectiveness and the feasibility of the improved algorithm is proved by the simulation experiment based on MATLAB.In the end of this paper, all of the work is conclude, the problems exist in the paper are appointed and the orientation of the future research is also advised.